Epic: PP-066 — 0.66 performance-parity release #2873

Activity

1. noahgift

   self-assigned this

   on Sep 5, 2026
2. noahgift

   added

   epicEpic — multi-story umbrellaEpic — multi-story umbrella

   on Sep 5, 2026
3. noahgift

   added a commit that references this issue on Sep 5, 2026

   docs(PP-066): 0.66 release spec v1.5 + driver prompt (PMAT-966, epic #…

   ...

   b132e2b
4. noahgift

   added sub-issues on Sep 5, 2026

   • Class: a gate that answers empty or degenerate input with PASS (or a traceback) instead of a named refusal #2876
   • Class: the required check dies of its environment (ENOENT, shared target dir, self-cancellation, runner $HOME, vanished binaries) and reports a code failure #2877
   • Class: universe authored, not derived — dark cargo targets, hand-enumerated CI lists and Makefile targets that advertise CI nothing invokes #2878
   • Class: ratchets that block edits rather than regressions — frozen files, line-keyed baselines and ledgers where honest work is the most expensive path #2879
5. noahgift commented on Sep 5, 2026

   

   noahgift

   on Sep 5, 2026

   ContributorAuthor

   More actions

   Proposed 0.66 row G-5 — for #2482 (Five failures surfaced by compiling 10 dark aprender-serve test target…), #2500 (aprender-test-lib fails strict clippy under its own non-default featur…), #2655 (bump-version.sh enumerates ONE excluded workspace by name instead of d…), #2674 (PARITY-007: the only apr-vs-llama.cpp bench in the tree is dark on thr…), #2751 (clippy reaches lib targets only: 1,782 tests/examples/benches/bins acr…), #2784 (driver_cuda_gguf.rs has 10 compile errors from struct changes: nothing…), #2808 (Five make targets advertise CI and no workflow invokes them — `coverag…)

   • A (acceptance command): scripts/check_target_universe.sh exit 0: for every workspace member (cargo metadata, root ∪ crates/facades) and every cargo target (lib, bins, tests, examples, benches) under the feature sets that compile it, the set of targets CI builds/lints/runs is derived from the workflows and compared against cargo metadata — any omission FAILs; cargo clippy --workspace --all-targets runs in reporting mode with a frozen per-crate shrink-only baseline; a Makefile target whose help says CI and that no workflow invokes FAILs
   • mutation → RED: drop one named --test from ci.yml l.408 while it exists in cargo metadata → RED; add # CI only to a make target no workflow calls → RED; restore the query_pre_attn_scalar omission in driver_cuda_gguf.rs → the cuda lane RED
   • expiry basis: anchored on master row 1's cuda-nightly.yml edit (2026-09-19) and C0's required-check work (S0-23) — the same files are opened
   • why in 0.66 scope (§2) with no existing row: §2's 0.66 lane includes guards; §9 targets '0 manifests classified by glob' (G-1) — the same principle over targets. Six open issues share the mechanism (driver_cuda_gguf.rs has 10 compile errors from struct changes: nothing builds aprender-serve --features cuda --tests, and cuda-nightly enumerates named targets so it never surfaces #2784 clippy reaches lib targets only: 1,782 tests/examples/benches/bins across 78 members are outside every lint gate #2751 Five failures surfaced by compiling 10 dark aprender-serve test targets (2 diagnosed, 3 open) #2482 aprender-test-lib fails strict clippy under its own non-default features — the trap that hid probar's dead code #2500 Five make targets advertise CI and no workflow invokes them — coverage-full says "CI only" #2808 bump-version.sh enumerates ONE excluded workspace by name instead of deriving the exclude list #2655; class parent filed by this triage): what is not on a written list is never built, linted or run, and R-0's own cuda-feature tests will land into the enumerated cuda lane
   • class parent: Class: universe authored, not derived — dark cargo targets, hand-enumerated CI lists and Makefile targets that advertise CI nothing invokes #2878 — universe authored, not derived — targets, scopes and lists enumerated by hand so what is not on the list is never built, linted or run

   No ticket created — the epic session owns pp-066-dag.yaml and pmat work add.

   PP-066 triage 2026-09-05 (issue #2873); ledger docs/audits/pp-066-triage-2026-09-05.md.
6. noahgift commented on Sep 5, 2026

   

   noahgift

   on Sep 5, 2026

   ContributorAuthor

   More actions

   Proposed 0.66 row I-24 — for #2735 (parity_block.py: a zero COMPARATOR band median is an unhandled traceba…)

   • A (acceptance command): python3 scripts/lib/parity_block.py --work … → rc=1 with 'band 8 comparator: non-positive rate is not a measurement' (named refusal); control fixture rc=0; zero-SUBJECT fixture keeps its existing named refusal (discrimination)
   • mutation → RED: restore the bare statistics.median(band["comparator"][metric]) division at scripts/lib/parity_block.py:571/652/680/720 → ZeroDivisionError traceback → RED
   • expiry basis: anchored on master row 15 (gx10 shakedown, derived) — the gx10 comparator is the lane that produced zero bands
   • why in 0.66 scope (§2) with no existing row: C4's post-publish parity attempt on four hosts runs this producer; an unhandled traceback on a zero comparator band is neither a refusal nor a verdict and leaves the C4 receipt without its block. No §5 card covers the producer's refusal polarity on the comparator side
   • class parent: Class: a gate that answers empty or degenerate input with PASS (or a traceback) instead of a named refusal #2876 — gate answers empty/degenerate input with a PASS or a crash instead of a named refusal

   No ticket created — the epic session owns pp-066-dag.yaml and pmat work add.

   PP-066 triage 2026-09-05 (issue #2873); ledger docs/audits/pp-066-triage-2026-09-05.md.
7. noahgift commented on Sep 5, 2026

   

   noahgift

   on Sep 5, 2026

   ContributorAuthor

   More actions

   Proposed 0.66 row I-25 — for #2756 (--workload is a free-text label decoupled from the prompts sent: a 1-p…)

   • A (acceptance command): cargo test -p apr-cli --test workload_corpus_binding — the receipt's workload is accepted only when corpus_sha256 in the receipt equals the sha256 of the prompts-w1.jsonl actually loaded and its _meta.corpus equals the label; apr test llm band --workload W1 --profile short is refused with one line
   • mutation → RED: restore the free-text Workload::from_str acceptance (label written without the digest) → RED
   • expiry basis: anchored on master row 15 (gx10 shakedown, derived) — the first CONFORMANT receipt must not carry an unfalsifiable label
   • why in 0.66 scope (§2) with no existing row: C1 requires a CONFORMANT W1 row; PP-22's join key includes the workload; today the label is a free string decoupled from the prompts sent (no corpus digest reaches the receipt: git grep corpus_sha256 origin/main -- crates/apr-cli/src → only apr-corpus-ingest.rs). No §5 card binds the label
   • class parent: Class: a gate that answers empty or degenerate input with PASS (or a traceback) instead of a named refusal #2876 — gate answers empty/degenerate input with a PASS or a crash instead of a named refusal

   No ticket created — the epic session owns pp-066-dag.yaml and pmat work add.

   PP-066 triage 2026-09-05 (issue #2873); ledger docs/audits/pp-066-triage-2026-09-05.md.
8. noahgift commented on Sep 5, 2026

   

   noahgift

   on Sep 5, 2026

   ContributorAuthor

   More actions

   Proposed 0.66 row G-6 — for paiml/paiml-mcp-agent-toolkit#1292 (pmat work add re-serialises all of docs/roadmaps/roadmap.yaml (2,531 u…)

   • A (acceptance command): scripts/check_roadmap_diff_additive.sh exit 0: the roadmap.yaml diff of a PR consists only of added entries (id set grows, no existing entry's lines change) — the driver's STEP 1b rule ('roadmap.yaml carries only the PMAT-966 hunk') as a guard wired beside check_roadmap_completion_is_cited.sh
   • mutation → RED: commit a pmat work add output un-trimmed (the 2,531-line re-serialisation of 6d9ba27) → RED; a trimmed 17-line hunk → GREEN
   • expiry basis: anchored on STEP 5 of the driver (every DAG row is minted with pmat work add) — before the first batch of PP-066 tickets is filed
   • why in 0.66 scope (§2) with no existing row: Every PP-066 ticket is minted with pmat work add (driver STEP 5); every such PR currently rewrites 714 unrelated entries, so two in-flight PP-066 PRs conflict by construction (The README contract count is a serialization point: any two contract-adding PRs collide by construction (4 occurrences today) #2630 class) and a reviewer cannot see the one real entry. The upstream fix is a paiml/pmat ticket; the guard is what keeps 0.66's PRs reviewable meanwhile

   No ticket created — the epic session owns pp-066-dag.yaml and pmat work add.

   PP-066 triage 2026-09-05 (issue #2873); ledger docs/audits/pp-066-triage-2026-09-05.md.
9. noahgift commented on Sep 5, 2026

   

   noahgift

   on Sep 5, 2026

   ContributorAuthor

   More actions

   Proposed 0.66 row G-7 — for #2775 (nextest reports a vanished test binary as a test failure, and cargo_cl…), #2822 (CI jobs fail with cargo exit 101: target/debug/deps files vanish MID-B…), #2828 (workspace-test and guard-runner-labels share one cargo target dir conc…), #2812 (Merge-queue runs are cancelled by their own concurrency key, and deque…), #2646 (Shim-check siting: required check depends on unversioned per-runner $H…)

   • A (acceptance command): scripts/cargo_classify.sh applied to the workspace-test job's captured log in ci.yml: an ENV verdict (double-spawn ENOENT, vanished .rlib/.o, ENOSPC, dep-info ENOENT) fails the job with ENV: in the check-run annotation and names no test; a CODE verdict names the test; bash scripts/cargo_classify.sh --self-test runs scripts/lib/cargo_failure_cases/*
   • mutation → RED: feed log_env_dep_info_contention.txt through the wired step → ENV, exit 1, no test named; feed log_code_test_failure.txt → CODE with the test name; swap the classifier for true → self-test RED
   • expiry basis: anchored on C0 (S0-23, 2026-09-19)
   • why in 0.66 scope (§2) with no existing row: the driver's rule 'a red ci/gate leg caused by CI jobs fail with cargo exit 101: target/debug/deps files vanish MID-BUILD (ENOENT, not ENOSPC — 884G free) #2822/workspace-test and guard-runner-labels share one cargo target dir concurrently with different feature sets — observed producing the ENOENT signature #2828 is recorded as infra' has no mechanism; the classifier exists with a fixture table and is wired to nothing (class parent: required check dies of its environment — CI jobs fail with cargo exit 101: target/debug/deps files vanish MID-BUILD (ENOENT, not ENOSPC — 884G free) #2822 workspace-test and guard-runner-labels share one cargo target dir concurrently with different feature sets — observed producing the ENOENT signature #2828 Merge-queue runs are cancelled by their own concurrency key, and dequeuing does not stop them — ~101 wasted runner-minutes measured #2812 Shim-check siting: required check depends on unversioned per-runner $HOME, while the dev box where the shim runs is never checked #2646 nextest reports a vanished test binary as a test failure, and cargo_classify.sh already knows that signature but is not wired to workspace-test #2775)
   • class parent: Class: the required check dies of its environment (ENOENT, shared target dir, self-cancellation, runner $HOME, vanished binaries) and reports a code failure #2877 — a required check dies of its environment and reports it as a code failure

   No ticket created — the epic session owns pp-066-dag.yaml and pmat work add.

   PP-066 triage 2026-09-05 (issue #2873); ledger docs/audits/pp-066-triage-2026-09-05.md.
10. noahgift

    added sub-issues on Sep 5, 2026

    • apr run: 7 defects incl. --stream emitting empty text and --trace-output always writing an empty event list #2378
    • APR-MONO consolidation debt: 126 verified findings — a member that will not compile, 9 orphans, guards scanning the wrong directory #2470
    • Four excluded crates cannot be loaded by any cargo command (metadata exits 101) #2492
11. 232 remaining items

    Load more

    Load more actions
12. noahgift

    added 5 commits that reference this issue on Sep 13, 2026

    docs(PP-066): 0.66 release spec v1.5 + driver prompt (PMAT-966, epic p…

    ...

    42be156

    docs(audits): PP-066 S0 discovery ledger — 23 premises measured (epic p…

    ...

    132ccda

    docs(audits): impl receipt for PMAT-742 — PARTIAL(andon) at K=150, ru…

    ...

    f6dcd10

    docs(PP-066): obligation DAG (87 rows, filed paiml#2882-paiml#2980), …

    ...

    aa5aa33

    docs(PP-066): session 2026-09-06 — SPEC-2.0 (v2.0, 18-row scope, clai…

    ...

    9465af2
13. noahgift

    closed this as completedin #3008on Sep 14, 2026
14. noahgift

    removed sub-issues on Oct 4, 2026

    • apr run: 7 defects incl. --stream emitting empty text and --trace-output always writing an empty event list #2378
    • apr gpu reports unified memory on GB10 but not on Apple silicon, and steers macOS users to --device cpu #2661
    • apr run --gpu on the 7B silently falls back to CPU after the F2 parity gate rejects at cosine 0.9278: any tok/s from that path is a CPU number with a GPU label #2790
    • Class: a gate that answers empty or degenerate input with PASS (or a traceback) instead of a named refusal #2876
    • BLOCKER: no W1 cell is producible on the comparator lane in either mode — streaming reports prompt_tokens=0, non-streaming reports 534 against a 512 ± 8 band #2817
    • apr mis-tokenizes <|im_end|> after sentence punctuation: 4 junk tokens are fed to the model instead of the turn delimiter #2815
    • Class: the required check dies of its environment (ENOENT, shared target dir, self-cancellation, runner $HOME, vanished binaries) and reports a code failure #2877
    • Class: ratchets that block edits rather than regressions — frozen files, line-keyed baselines and ledgers where honest work is the most expensive path #2879
    • Class: universe authored, not derived — dark cargo targets, hand-enumerated CI lists and Makefile targets that advertise CI nothing invokes #2878